#6635 [Benchmark] RTX PRO 4500, EPYC 73F3, 256GB

open · @miyconst · 0 コメント · GitHub で見る

BenchmarksAMD / HIPNVIDIA / CUDAModels & quantsWindows

本文

# Strata standard benchmark: NVIDIA RTX PRO 4500 Blackwell

Protocol `strata-qwen38-thinking-v1-draft.1` (draft), run `20261122T160042Z-78e01cb2`, started 2026-11-22T16:00:42Z. **Status: complete.**

Written by `strata bench run`. The machine-readable result is [results.json](results.json); every request, its raw usage and timings and the client event offsets are in [outputs.jsonl](outputs.jsonl).

## Hardware and software

- GPU: NVIDIA RTX PRO 4500 Blackwell (31.9 GiB dedicated)
- CPU: AMD EPYC 73F3 16-Core Processor; RAM: 255.8 GiB
- Storage of the model files: ssd SAMSUNG MZVLB512HAJQ-000H2 (nvme)
- OS: Windows 10.0.26100; drivers: gpu-driver 621.24; runtime: cuda 13.5
- Strata commit: `26350e4387b8019b0f6825ba78b87eb6360159e7`; engine 0.1.65; release build
- Background work: not measured

## Model and configuration

- Model: qwen3.8-flash-next-iq3_s; quantization IQ3_S; revision not measured; source revision not measured
- Files: `Qwen3.8-Flash-Next-GSQ-RCO-IQ3_S-00001-of-00002.gguf` `4c1eb2ceb4915e11...`; `Qwen3.8-Flash-Next-GSQ-RCO-IQ3_S-00002-of-00002.gguf` `316b46f3a2dbd68c...`
- Track: **custom** (no notes given by the operator)
- Context 131072 tokens; server capacity 1 slot(s)
- Sampling: temperature 1.0, top_p 0.95, top_k 20, min_p 0.0, presence_penalty 0.0, repetition_penalty 1.0, reasoning_effort xhigh; thinking on, preserved
- Engine options: `config.json`; chat template: `provenance/chat_template.jinja`

## Method

The frozen requests of `bench/prompts/standard-v1` (corpus `e75eff8b9feea04f...`) were sent over loopback HTTP, streaming, one at a time, in this order: 1 warm-up (excluded), 3 repetitions of the short, 4K and 16K prompts, and the twelve sanity checks. Prompt and decode speed are the engine's own `timings`; time to first token is the client's monotonic clock from dispatch to the first nonempty reasoning, answer or tool-call delta. No request was retried. Time after the model loaded: 1 min 6 s.

## Results 128k Context

| Case | Prompt tok/s | Decode tok/s | TTFT s | First answer s | Total s | Capped |
| --- | --- | --- | --- | --- | --- | --- |
| short | n/a | 154.3 (151.4-155.1) n=3 | 0.41 (0.28-0.42) n=3 | 0.92 (0.88-0.98) n=3 | 2.7 (2.2-3.0) n=3 | 0/3 |
| prompt-4k | 3186.5 (3168.9-3225.9) n=3 | 146.2 (131.2-156.9) n=3 | 1.30 (1.28-1.31) n=3 | 3.45 (3.11-4.59) n=3 | 4.8 (4.5-5.2) n=3 | 3/3 |
| prompt-16k | 4361.3 (4306.5-4422.9) n=3 | 124.7 (120.8-132.7) n=3 | 3.80 (3.73-3.84) n=3 | 7.22 (7.22-7.22) n=1 | 7.8 (7.6-8.1) n=3 | 3/3 |

## Results 256k Context

| Case | Prompt tok/s | Decode tok/s | TTFT s | First answer s | Total s | Capped |
| --- | --- | --- | --- | --- | --- | --- |
| short | n/a | 151.6 (143.9-159.3) n=3 | 0.42 (0.28-0.42) n=3 | 1.22 (0.97-1.52) n=3 | 2.9 (2.8-3.3) n=3 | 0/3 |
| prompt-4k | 3166.2 (3164.0-3177.1) n=3 | 146.6 (143.2-158.9) n=3 | 1.31 (1.31-1.31) n=3 | 3.33 (3.23-4.03) n=3 | 4.8 (4.5-4.9) n=3 | 3/3 |
| prompt-16k | 4374.8 (4329.8-4424.7) n=3 | 133.2 (129.7-142.9) n=3 | 3.77 (3.73-3.81) n=3 | 6.66 (5.99-7.33) n=2 | 7.6 (7.3-7.7) n=3 | 3/3 |

[2026-11-22-standard-rtx-pro-4500-blackwell-78e01cb2.zip](https://github.com/user-attachments/files/35238113/2026-11-22-standard-rtx-pro-4500-blackwell-78e01cb2.zip)
[2026-11-22-standard-rtx-pro-4500-blackwell-2607ffad.zip](https://github.com/user-attachments/files/35238112/2026-11-22-standard-rtx-pro-4500-blackwell-2607ffad.zip)

関連リンク

インストール・モデル・リリースへのサイト内リンク。